News
Google DeepMind's Sierra Leone experiment: students with AI tutor gained over a year of learning in eight weeks
A randomized study by Google DeepMind covered nearly 1,800 students in Sierra Leone using Guided Learning, a Gemini-based tutoring tool - test scores improved enough to correspond to 1.2 to 1.7 years of typical learning progress.
Contents
Google DeepMind has published the results of an eight-week, pre-registered randomized study in which nearly 1,800 seventh and eighth grade students across 48 classrooms in Sierra Leone used Guided Learning, an educational tool built on Gemini. The effect: math test scores rose by 0.26 standard deviations, which under the study's methodology corresponds to between 1.2 and 1.7 years of typical learning progress in low- and middle-income countries.
How Guided Learning works
Guided Learning isn't just a regular chat with Gemini, but a version of the model rebuilt specifically for teaching. Instead of handing over ready-made solutions, the tool is designed to walk students through a problem using guiding questions. Study data show that in the vast majority of conversations, students arrived at understanding on their own, with Gemini providing a direct answer in only about 2 percent of messages.
The study was led by Irina Jurenka, a research director at Google DeepMind, who herself cautioned that the headline figure of "a year of learning in eight weeks" is an estimate based on limited data, not a hard certainty. Asked directly about the system's reliability, she said that since it's a language model, no one can guarantee its answers are 100 percent accurate - the team simply tried to steer the system as closely as possible toward pedagogically sound responses.
Who benefited, and who didn't
69 percent of students reached the recommended usage threshold of the tool (12 hours), with average engagement at 15 hours per student - well above the typical 5 percent voluntary adoption rate for educational tools. Students who met the recommended usage threshold saw a bigger jump, 0.38 standard deviations, which moved average students into the top third of their class.
The study also revealed a significant equity problem: the students who gained the most were those who already performed best in math beforehand. Jurenka acknowledged this risks widening the gap between students and said the team plans further work on pedagogical approaches better suited to weaker students. The final test was half made up of material not previously covered in class, meant to check whether the learning actually generalized - according to the team, signs of broader improvement did appear.
The other side of the project: Italy
Google also described a parallel program at the Don Bosco school network in northern Italy, covering more than 560 classes, 700 teachers and 9,000 students from primary school through vocational training, run on Gemini for Education. Between 80 and 99 percent of students mastered the planned skills in individual lessons, and teachers reported a 70 percent cut in time spent on administrative tasks, which they redirected toward one-on-one mentoring and emotional support for students.
Google also announced it would expand its AI Educator Series program in India with localization into six languages (Assamese, Hindi, Marathi, Telugu, Odia, Punjabi), along with a partnership with the African Union Commission to support the rollout of Gemini for Education and NotebookLM at universities across 55 member states.
What this means for Polish schools
In Poland, the debate about AI in the classroom currently revolves mostly around what teachers and students shouldn't do, rather than around concrete rollouts at the scale of thousands of students. The results from Sierra Leone and Italy are among the first publicly documented, controlled studies showing a measurable effect from an AI tool designed specifically for teaching rather than for generating ready-made answers - a key distinction given concerns about chatbots doing students' homework for them.
DeepMind published its methodology, its research decision-making process, and teacher training materials to allow the experiment to be independently replicated. For Poland's education system, where the Ministry of Education (MEN) and regional school boards (kuratoria) are still working out guidelines on AI in schools, such a well-documented pilot model could serve as a reference point before any similar tool reaches Polish classrooms at a larger scale.
Sources: Google Tested Its AI Tutor In Real Classrooms. It Worked (forbes.com), Measuring the impact of AI on teaching and learning (blog.google).


